Portfolio Construction Using Argumentation and Hybrid ...
Transcript of Portfolio Construction Using Argumentation and Hybrid ...
1
Portfolio Construction Using Argumentation and
Hybrid Evolutionary Forecasting Algorithms
Nikolaos Spanoudakis*
Applied Mathematics and
Computers Laboratory,
Department of Sciences,
Technical University of Crete,
University Campus, 73100,
Kounoupidiana, Greece.
Konstantina Pendaraki
Department of Business
Administration of Food and
Agricultural Enterprises,
University of Ioannina,
G. Seferi 2, 30100 Agrinio,
Greece.
Grigorios Beligiannis
Department of Business
Administration of Food and
Agricultural Enterprises,
University of Ioannina,
G. Seferi 2, 30100 Agrinio,
Greece.
Abstract. In the current contribution, an application for constructing mutual fund portfolios is presented. This approach
comprises of several Intelligent Methods, namely an argumentation based decision making framework and a hybrid evolutionary
forecasting algorithm which combines Genetic Algorithms (GA), MultiModel Partitioning (MMP) theory and Extended Kalman
Filters (EKF). Specifically, the argumentation framework is employed in order to develop mutual funds performance models and
to select a small set of mutual funds, which will compose the final portfolio. On the other hand, the hybrid evolutionary
forecasting algorithm is applied in order to forecast the market status (inflating or deflating) for the next investment period. The
knowledge engineering approach and application development steps are also presented and discussed.
Keywords. Portfolio Construction, Mutual Funds, Intelligent Methods, Argumentation, Forecasting, Hybrid Evolutionary
Algorithms
* Corresponding author.
Tel.: +30 28210 37744
Fax: +30 28210 37842
E-mail: [email protected]
2
1 INTRODUCTION
Portfolio management [21] is concerned with constructing a portfolio of securities (e.g., stock, bonds, mutual funds
[32], etc.) that maximizes the investor’s utility. In a previous study [34], we constructed mutual fund (MF) portfolios
using an argumentation based decision making framework. We developed rules that characterize the market and
different investor types policies using evaluation criteria of fund performance and risk. We also defined strategies for
resolving conflicts over these rules. Furthermore, the developed application can be used for a set of different
investment policy scenarios and supports the investor/portfolio manager in composing efficient MF portfolios that
meet his investment preferences. Traditional portfolio theories ([21], [30], [31]) have been based on unidimensional
approaches that do not fit to the multidimensional nature of risk ([8]), and they do not capture the complexity
presented in the data set. In [34], this troublesome situation was resolved by the high level of adaptability in the
decisions of the portfolio manager or investor, when his environment is changing and the characteristics of the funds
are multidimensional, that was demonstrated by the use of argumentation.
Argumentation was selected among a number of different approaches as it: (a) allows for decision making using
conflicting knowledge, (b) allows defining non-static priorities between arguments, and (c) the modularity of its
representation allows for the easy incorporation of views of different experts. Traditional approaches such as
statistical methods need to make strict statistical hypothesis, multi-criteria analysis methods need significantly more
effort from experts, and neural networks require increased computational effort and are characterized by inability to
provide explanations for the results.
Our study showed that when taking into account the market context, the results would be better if we could forecast
the status of the market of the following investment period. In order to achieve this goal we employed a hybrid
system that combines Genetic Algorithms (GA), MultiModel Partitioning (MMP) theory and the Extended Kalman
Filter (EKF). A general description of this algorithm and its application in linear and non-linear data is discussed in
[4], while the specific version used in this contribution is presented in [1], where its successful application to non-
linear data is also presented. This algorithm captured our attention because it had been successfully used in the past
for accurately predicting the evolution of stock values in the Greek market (its application on economic data is
presented in [4]). Moreover, there is a lot of work on hybrid evolutionary algorithms and their application on many
difficult problems has shown very promising results [9]. The problem of predicting the behavior of the financial
3
market is an open problem and many solutions have been proposed ([21], [22], [25], [34]). However, there isn't any
known algorithm able to identify effectively all kinds of behaviors. Also, many traditional methods have been
applied to the same problem and the results obtained were not very satisfactory. There are two main difficulties in
this problem, firstly the search space is huge and, secondly, it comprises of many local optima. We chose to apply a
hybrid evolutionary system to this problem because Evolutionary Algorithms comprise a global forecasting
technique, in the sense that a single formula is sought that allows forecasts of future entries in any series generated by
the process—starting at any point in time. In this way, they perform well in cases where the search space is huge
and/or comprises of many local optima.
In this contribution, we present the new portfolio construction application resulting from the combination of
argumentation with hybrid evolutionary systems along with the obtained results.
The rest of the paper is organized as follows: Section two presents an overview of the concepts and application
domain knowledge along with a presentation of the data set that we use. Section three outlines the main features of
the proposed argumentation based decision-making framework and the developed argumentation theory. The
forecasting hybrid evolutionary system is presented in section four, followed by section five, which presents the
developed application and discusses the obtained empirical results. Finally, section six summarizes the main findings
of this research and discusses its added value compared to different approaches in the past.
2 DOMAIN KNOWLEDGE
This section firstly presents the sources of the data that we used for our study. Subsequently, it describes the criteria
(or variables) used for creating portfolios and the knowledge on how to use these criteria in order to construct a
portfolio.
2.1 The Sample Data Set
The sample used in this study is provided by the Association of Greek Institutional Investors and consists of daily
data of domestic equity mutual funds (MFs) over the period January 2000 to December 2007. Daily returns for all
domestic equity MFs are examined for this eight-year period. At the end of 2007, the sample consisted of 80
domestic equity MFs. Nevertheless, not all MFs have been in operation for the whole eight-year time period of the
4
analysis. Therefore, in order to eliminate the effect that could be caused by the fact that not all MFs were in
operation for the same period, it was decided to consider only the ones for which complete data were available for
each sub-period that the calculations were made.
For the application of the proposed methodological framework, further information is derived from the official web
pages of the Athens Stock Exchange (www.ase.gr) and the Alpha Bank of Greece (www.alpha.gr), regarding the
return of the Athens stock exchange market and the return of the three-month Treasury bill respectively. This
information is very important as the variations of the returns in the ASE-GI, (Athens Stock Exchange-General Index)
are expressed through the variation of the prices of the stocks taking also into account the fluctuations and the risk of
the financial environment. Furthermore, the three-month Treasury bill is a risk-free asset, and constitutes the
benchmark where the return of a MF or another risky investment is compared with.
2.2 Evaluation criteria
The proposed framework is based on five fundamental criteria, which represent basic characteristics of the examined
funds. These characteristics of the examined funds refer to different performance and risk variables. The return
variables are measuring the expected outcome of the investment in the MFs, while the risk variables are measuring
the uncertainty about the outcome of the investment. These criteria are the following: (1) the return of the funds, (2)
the standard deviation of the returns, (3) the beta coefficient, (4) the Sharpe index, and (5) the Treynor index. These
variables are frequently used in portfolio management research area (see Table 1).
Table 1 Representative bibliography references of the examined criteria
Criteria References
(1) Fund’s return [6], [33], etc.
(2) Standard deviation of the returns [19], [13], [12], etc.
(3) beta coefficient [14], [33], etc.
(4) Sharpe index [32], [16], etc.
(5) Treynor index [36], [29], etc.
5
A brief description of these criteria follows. The return of the funds is the actual value of return of an investment
defined by the difference between the nominal return and the rate of inflation. This variable is based on the net price
of a fund. At this point, it is very important to mention that transaction costs such as management commission are
included in the net price. Frond-end commission and redemption commission fluctuate depending on the MF class
and in most cases are very low. The standard deviation is used to measure the variability of the fund’s daily returns,
thus representing the total risk of the fund. The beta coefficient (β) is a measure of fund’s risk in relation to the
capital risk. The Sharpe index [32] is a useful measure of performance and is used to measure the expected return of
a fund per unit of risk, defined by the standard deviation. The Treynor index [36] is similar to the Sharpe index
except that performance is measured as the risk premium per unit of systematic (beta coefficient) and not of total
risk. The exact formulas by which these criteria are calculated are presented on Table 2.
On the basis of the argumentation framework for the selection of a small set of MF, which will compose the final
multi-portfolios, the examined funds are clustered in three groups for each criterion for each year. For example, we
have funds with high, medium and low performance for the return on funds variable, the same for the other criteria.
The aforementioned performance and risk variables visualize the characteristics of the capital market (bull or bear)
and the type of the investor according to his investment policy (aggressive or moderate). Further information is
represented through variables that describe the general conditions of the market and the investor policy (selection of
portfolios with high performance per unit of risk).
The general conditions of the market are characterized through the development of funds which have high
performance levels (high return on funds). Regarding the market context, in a bull market, funds are selected if they
have high systematic or total risk. On the other hand, in a bear market, we select funds with low systematic and total
risk. An aggressive investor is placing his capital upon funds with high performance and high systematic risk.
Accordingly, a moderate investor selects funds with high performance and low or medium systematic risk. Some
types of investors select portfolios with high performance per unit of risk. Such portfolios are characterized by high
Sharpe ratio and high Treynor ratio.
6
Table 2. Criteria used
Criteria Formulas Variables definitions
(1) Fund’s return
1
1
t
ttpt
NAV
NAVDISTNAVR Rpt = return of MF in period t,
NAVt = closing net asset value of the fund on the last
trading day of the period t,
NAVt-1 = closing net asset value of the fund on the last
trading day of the period t-1,
DISTt = income and capital distributions (dividend of
the fund) taken during period t
(2) Standard
deviation of
the returns
2)(1
ptpt RRT
= standard deviation of MF in period t,
ptR = average return in period t,
T = number of observation (days) in the period for
which the σ is being calculated
(3) beta
coefficient
cov (Rpt, RMt) / var (RMt) RMt = return of market portfolio (Athens Stock
Exchange) in period t,
cov (Rpt, RMt) = covariance of daily return of MF with
market portfolio,
var (RMt) = variance of daily return of market portfolio
(4) Sharpe index (Rpt - Rft) / σ Rft = return of treasure bill in period t
(5) Treynor index (Rpt - Rft) / β
3 ARGUMENTATION-BASED DECISION MAKING
In this section we firstly present the argumentation framework that we used and then we describe the domain
knowledge modeling.
7
3.1 The Argumentation Framework
Autonomous agents, be they artificial or human, need to make decisions under complex preference policies that take
into account different factors. In general, these policies have a dynamic nature and are influenced by the particular
state of the environment in which the agent finds himself. The agent's decision process needs to be able to synthesize
together different aspects of his preference policy and to adapt to new input from the current environment. Such
agents are the mutual fund managers.
In order to address requirements like the above, Kakas and Moraitis ([17]) proposed an argumentation based
framework to support an agent's self deliberation process for drawing conclusions under a given policy.
Argumentation can be abstractly defined as the principled interaction of different, potentially conflicting arguments,
for the sake of arriving at a consistent conclusion (see e.g. [27]). The nature of the “conclusion” can be anything,
ranging from a proposition to believe, to a goal to try to achieve, to a value to try to promote. Perhaps the most
crucial aspect of argumentation is the interaction between arguments. This means that argumentation can give us
means for allowing an agent to reconcile conflicting information within itself, for reconciling its informational state
with new perceptions from the environment, and for reconciling conflicting information between multiple agents
through communication. A single agent may use argumentation techniques to perform its individual reasoning
because it needs to make decisions under complex preferences policies, in a highly dynamic environment (see e.g.
[17]). This is the case used in this research. In the following paragraphs we describe the theoretical framework that
we adopted:
Definition 1. A theory is a pair (T, P) whose sentences are formulae in the background monotonic logic (L, ⊢) of
the form L←L1,…,Ln, where L, L1, …, Ln are positive or negative ground literals. For rules in P the head L refers to
an (irreflexive) higher priority relation, i.e. L has the general form L = h_p(rule1, rule2). The derivability relation, ⊢,
of the background logic is given by the simple inference rule of modus ponens.
An argument for a literal L in a theory (T, P) is any subset, T, of this theory that derives L, T ⊢L, under the
background logic. A part of the theory T0 ⊂ T, is the background theory that is considered as a non defeasible part
(the indisputable facts).
8
An argument attacks (or is a counter argument to) another when they derive a contrary conclusion. These are
conflicting arguments. A conflicting argument (from T) is admissible if it counter-attacks all the arguments that
attack it. It counter-attacks an argument if it takes along priority arguments (from P) and makes itself at least as
strong as the counter-argument (we omit the relevant definitions from [17] due to limited space).
Definition 2. An agent’s argumentative policy theory is a theory T = ((T, T0), PR, PC) where T contains the
argument rules in the form of definite Horn logic rules, PR contains priority rules which are also definite Horn rules
with head h_p(r1, r2) s.t. r1, r2 ∈ T and all rules in PC are also priority rules with head h_p(R1, R2) s.t. R1, R2 ∈ PR
∪ PC. T0 contains auxiliary rules of the agent’s background knowledge.
Thus, in defining the decision maker’s theory we specify three levels. The first level (T) defines the (background
theory) rules that refer directly to the subject domain, called the Object-level Decision Rules. In the second level we
have the rules that define priorities over the first level rules for each role that the agent can assume or context that he
can be in (including a default context). Finally, the third level rules define priorities over the rules of the previous
level (which context is more important) but also over the rules of this level in order to define specific contexts, where
priorities change again.
3.2 The Decision Maker’s Argumentation Theory
Using the presented argumentation framework, we transformed the criteria for all MFs and experts knowledge (§2)
to background theory (facts) and rules of the first and second level. Then, we defined the strategies (or specific
contexts) in the third level rules.
The goal of the knowledge base is to select some MFs in order to construct our portfolio. Therefore, our rules have
as their head the predicate selectFund/1 and its negation. We write rules supporting it or its negation and use
argumentation for resolving conflicts. We introduce the hasInvestPolicy/2, preference/1 and market/1 predicates for
defining the different contexts and roles. For example, John, an aggressive investor is expressed with the predicate
hasInvestPolicy(john, aggressive).
9
The knowledge base facts are the performance and risk variables values for each MF, the thresholds for each group
of values for each year and the above mentioned predicates characterizing the investor and the market. The following
rules are an example of the object-level rules (level 1 rules of the framework - T):
r1(Fund): selectFund(Fund) ← highR(Fund)
r2(Fund): ¬selectFund(Fund) ← highB(Fund)
The highR predicate denotes the classification of the MF as a high return fund and the highB predicate denotes the
classification of the MF as a high risk fund. Thus, the r1 rule states that a high performance fund should be selected,
while the r2 rule states that a high risk fund should not be selected. Such rules are created for the three groups of our
performance and risk criteria.
Then, in the second level we assign priorities over the object level rules. The PR are the default context rules or level
2 rules. These rules are added by experts and express their preferences in the form of priorities between the object
level rules that should take place within defined contexts and roles. For example, the level 1 rules with signatures r1
and r2 are conflicting. In the default context the first one has priority, while the bear market context reverses this
priority:
R1: h_p(r1(Fund),r2(Fund)) ← true
R2: h_p(r2(Fund),r1(Fund)) ← market(bear)
Rule R1 defines the priorities set for the default context, i.e. an investor selects a fund that has high return on
investment (RoI) even if it has high risk. Rule R2 defines the default context for the bear market context (within
which, the fund selection process is cautious and does not select a high RoI fund if it has high risk).
Finally, in PC (level 3 rules) the decision maker defines his strategy and policy for integrating the different roles and
contexts rules. When combining the Aggressive investor role and bear market context, for example, the final
10
portfolio is their union except that the aggressive investor now would accept to select high and medium risk MFs
(instead of only high). The decision maker’s strategy sets preference rules between the rules of the previous level but
also between rules at this level. Relating to the level 2 priorities, the bear market context’s priority of not buying a
high risk MF, even if it has a high return, is set at higher priority than that of the general context. Then, the specific
context of an aggressive investor in a bear market defines that the bear market context preference is inverted. See the
relevant priority rules:
C1: h_p(R2, R1) ← true
C2: h_p(R1, R2) ← hasInvestPolicy(Investor, aggressive).
C3: h_p(C2, C1) ← true
Thus, an aggressive investor in a bear market context would continue selecting high risk funds. In the latter case, the
argument r1 takes along the priority arguments R1, C2 and C3 and becomes stronger (is the only admissible one) than
the conflicting r2 argument that can only take along the R2 and C1 priority arguments. Thus, the selectFund(Fund)
predicate is true and the fund is inserted in the portfolio.
The problem with the above rules is that the facts market(bear) or (exclusive) market(bull) could not be safely
determined for the next investment period. In the application version presented in [34] it was just assumed to remain
the same as at the time of the investment. This strategy, however, produced quite poor results for this context if it
should change in the next period.
4 FORECASTING THE STATUS OF THE FINANCIAL MARKET
One of the most prominent issues in the field of signal processing is the adaptive filtering problem, with unknown
time-invariant or time-varying parameters. Selecting the correct order and estimating the parameters of a system
model is a fundamental issue in linear and nonlinear prediction and system identification. The problem of fitting an
AutoRegressive Moving Average model with eXogenous input (ARMAX) or a Nonlinear AutoRegressive Moving
11
Average model with eXogenous input (NARMAX) to a given time series has attracted much attention because it
arises in a large variety of applications, such as time series prediction in economic and biomedical data, adaptive
control, speech analysis and synthesis, neural networks, radar and sonar, fuzzy systems, and wavelets [15].
The forecasting algorithm used in this contribution is a generic applied evolutionary hybrid technique, which
combines the effectiveness of adaptive multimodel partitioning filters and GAs’ robustness [1]. This method has
been first presented in [18]. Specifically, the a posteriori probability that a specific model, of a bank of the
conditional models, is the true model, can be used as fitness function for the GA. In this way, the algorithm identifies
the true model even in the case where it is not included in the filters’ bank. It is clear that the filter’s performance is
considerably improved through the evolution of the population of the filters’ bank, since the algorithm can search the
whole parameter space. The proposed hybrid evolutionary algorithm can be applied to linear and nonlinear data; is
not restricted to the Gaussian case; does not require any knowledge of the model switching law; is practically
implementable, computationally efficient and applicable to online/adaptive operation; and, finally, exhibits very
satisfactory performance as indicated by simulation experiments [4]. The structure of the hybrid evolutionary system
used is depicted in Fig. 1.
The representation used for the genomes of the population of the GA is the following. We use a mapping that
transforms a fixed dimensional internal representation to variable dimensional problem instances. Each genome
consists of a vector x of real values xi , i = 1, ..., k, and a bit string b of binary digits bi{0,1}, i = 1, ..., k. Real
values are summed up as long as the corresponding bits are equal. Obviously, k is an upper bound for the dimension
of the resulting parameter vector. We use the first k/3 real values for the autoreggressive part, the second k/3 real
values for the moving average part, and the last k/3 real values for the exogenous input part. An example of this
mapping is presented in Fig. 2. For a more detailed description of this mapping refer to [4]. In this specific
implementation the number of genes comprising each genome of the GA equals to 12. Subsequently and according to
our model, the first 4 real values are used for the autoreggressive part, the second 4 real values are used for the
moving average part, and the last 4 real values are used for the exogenous input part.
At first, an initial population of m genomes is created at random (each genome consists of a vector of real values and
a bit string). In this specific implementation the number of possible solutions that are evolved, that is the population
size of the GA, equals to 20. As stated before, each vector of real values represents a possible value of the
12
NARMAX model order and its parameters. For each such population we apply an MMAF with EKFs and have as
result the model-conditional probability density function (pdf) of each candidate model. This pdf is the fitness of
each candidate model, namely the fitness of each genome of the population (Fig. 3).
Fig. 1: The structure of the hybrid evolutionary system used for forecasting.
Fig. 2: Mapping from a fixed dimensional internal representation to a variable length NARMAX parameter vector. The
resulting order is n(p, q, r) = (4, 3, 2).
13
Fig. 3: The fitness of each candidate model is the model conditional pdf (m is the number of the extended Kalman filters in
the multimodel adaptive filter).
The reproduction operator we decided to use is the classic biased roulette wheel selection according to the fitness
function value of each possible model order [22]. As far as crossover is concerned, we use the one-point crossover
operator for the binary strings and the uniform crossover operator for the real values [22]. For both crossover
operators a probability equal to 0.9 was used. Finally, we use the flip mutation operator for the binary strings and the
Gaussian mutation operator for the real values [22]. For both mutation operators a probability equal to 0.1 was used.
Every new generation of possible solutions iterates the same process as the old ones and all this process may be
repeated as many generations as we desire or till the fitness function has value 1 (one) which is the maximum value it
is able to have as a probability. For a more detailed description of this hybrid evolutionary system refer to [4].
In this contribution we apply a slightly different approach compared to the one presented in [4]. In [4], at the
algorithm’s step where the value of the estimation (output) x of each filter is calculated, the past values of x that are
used in order to estimate the next value of x are always taken from the estimation file (the file of all past values of x
that have been estimated by the algorithm till this point). All these values are used in each generation in order to
estimate the next value of the estimation (output) vector x. The method presented in this contribution uses a different
approach in order to estimate x. At the algorithm’s step where the value of x for each filter is calculated, the past
values of x that are used in order to estimate the next value of x are smaller than the total length of the time series that
has been estimated till this point. The length of past values used in each generation in order to estimate the next value
of x equals to n/2, where n is the total length of the time series to be estimated. Every new value of x, estimated by
the algorithm, is added to this time series of length n/2 and the oldest one is removed in order this time series to
14
sustain a length of n/2. The value of n/2 was not selected arbitrarily. We have conducted exhaustive experiments
using many different values. The value of n/2, that has been finally selected, was the most effective one, that is, the
one that resulted in the best prediction results. In this specific implementation the value of parameter n equals 30,
that is 30 semesters are used as past values for forecasting, starting from the first semester of 1985. Of course, as
stated before, every new value of x (status of the financial market for a specific semester), estimated by the
algorithm, is added to this time series of length 30 and the oldest one is removed in order this time series to sustain a
length of 30.
Thus, the hybrid evolutionary system presented in Fig. 1 is used in order to forecast the behavior of the financial
market in relation to its current status. The market is characterized as bull market if it is forecasted to rise in the next
semester, or as bear market if it is forecasted to fall. We used the percentage of the return of the Athens Stock
Exchange index for each semester (in relation to the previous semester) starting from year 1985 to the years of our
sample data (2000 to 2007). Our algorithm indicates a bull market if this percentage is forecasted to be positive or a
bear market if it is forecasted to be negative. The algorithm performed very well considering that it could forecast the
next semester market behavior with a success rate of 88.24% (15 out of 17 right predictions), as presented in Table
3.
One can argue, after studying the results presented in Table 3, that the forecasted values of change are far different
from the actual values of the change of the financial market index. However, the algorithm is not applied in order to
predict specific values of the financial market index. It is used, as stated above, in order to predict the next semester
financial market behavior, which is the behavior of the financial market in relation to its current status (whether the
financial market index of the current semester is higher or lower compared to the financial market index of the
previous semester). Results presented in Table 3 demonstrate that the proposed forecasting algorithm achieves very
satisfactory results in doing this.
15
5 THE PORTFOLIO CONSTRUCTION APPLICATION
In this section we firstly present the system architecture, i.e. the combination method for the argumentation decision
making sub-system and the hybrid forecasting sub-system that resulted in a coherent application. Then we present the
results of this combination.
Table 3: Results obtained after applying the presented forecasting algorithm in order to forecast the sign (positive or
negative) of the return of the Athens Stock Exchange index for each semester (the rows with grey background indicate the
failed forecasts).
Semester RASE change (%) Forecasted value
1st sem 2000 -26.751 5.552
2nd sem 2000 -15.706 -4.006
1st sem 2001 -19.112 -2.409
2nd sem 2001 -5.267 -2.989
1st sem 2002 -14.822 -0.826
2nd sem 2002 -21.206 -2.334
1st sem 2003 6.468 -3.412
2nd sem 2003 21.190 1.025
1st sem 2004 1.535 3.391
2nd sem 2004 19.219 6.656
1st sem 2005 8.357 3.067
2nd sem 2005 19.204 1.343
1st sem 2006 0.831 3.118
2nd sem 2006 18.984 1.091
1st sem 2007 8.440 0.0125
2nd sem 2007 6.561 0.3002
1st sem 2008 -33.946 -0.002
16
5.1 System Architecture
The portfolio generation application is a Java program creating a human-machine interface and managing its
modules, namely the decision making module, which is a prolog rule base (executed in SWI-prolog†) using the
Gorgias‡ framework, and the forecasting module, which is a Matlab
§ implementation of the forecasting hybrid system
(see Fig. 4).
Portfolio Construction
Application
Java application
Decision Making Module
SWI-Prolog, Gorgias
Forecasting Module
Matlab
Facts
MySQL
execJPL
JDBC
SQL
Fig. 4: System Architecture.
The application connects to the SWI-Prolog module using the provided Java interface (JPL) that allows for inserting
facts to an existing rule-base and running it for reaching goals. The goals can be captured and returned to the Java
program. The application connects to Matlab by executing it in a system shell. The matlab program writes the results
of the algorithm to a MySQL**
database using SQL (Structured Query Language). The application first executes the
forecasting module, then updates the database, using JDBC (Java DataBase Connectivity interface) technology, with
the investor profile (selected roles) and, finally, queries the decision making module setting as goal the funds to
select for participation in the final portfolio. Thus, after the execution of the forecasting module the predicate
† SWI-Prolog offers a comprehensive Free Software Prolog environment, http://www.swi-prolog.org
‡ Gorgias is an open source general argumentation framework that combines the ideas of preference reasoning and abduction,
http://www.cs.ucy.ac.cy/~nkd/gorgias/
§ MATLAB® is a high-level language and interactive environment for performing computationally intensive tasks,
http://www.mathworks. com/products/matlab
** MySQL is an open source database, http://www.mysql.com
17
market/1 is determined as bull or bear and inserted as a fact in the rule base before the decision making process is
launched. The reader can see in Fig. 5 a screenshot of the integrated system.
Fig. 5: A screenshot for portfolio generation for a scenario of a moderate investor in a bull market context.
5.2 System Evaluation
For evaluating our results we defined scenarios for all years for which we had available data (2000-2007) and for all
combinations of contexts. That resulted to five simple scenarios (general, market, aggressive investor, moderate
investor and high performance contexts) and five specific context scenarios (the two investor types combined with
the market status, plus the two investor types combined with the high performance option, plus the market status
combined with the high performance option) all run for eight years each. Each one of the examined scenarios refers
to different investment choices and leads to the selection of different number and combinations of MFs.
Firstly, we use a limited range of these data (years 2000-2005) for comparing the hybrid approach presented in this
paper to the previous argumentation-based approach ([34]). In Table 4 the reader can inspect the average return on
investment (RoI) for the first six years (2000-2005) for all different contexts. The reader should notice that the table
contains two RoI columns, the first (“Previous RoI”) depicts the results before changing the system as they appeared
in [34]. The second presents the results of upgrading the application by combining it with the hybrid evolutionary
forecasting sub-system and by fixing the selected funds participation to the final portfolio. The latter modification is
18
out of the scope of this paper but the reader can clearly see that it has greatly influenced the performance of all
scenarios. Thus, Table 4, shows the added value of this contribution as the market context has become the most
profitable in the “New RoI” column (8.17% RoI), while in the “Previous RoI” column it was one of the worst cases
(3.72% RoI). Consequently, the specific contexts containing the market context have better results.
Table 4. Average RoI for six years. The New RoI column shows the gains after the evolutionary hybrid forecasting
system’s integration
Context type Context Previous RoI (%) New RoI (%)
simple general 3.53 6.86
role aggressive 2.65 7.38
role moderate 4.02 6.09
context market 3.72 8.17
role high performance 4.98 7.16
specific context aggressive – market 3.56 7.92
specific context moderate – market 4.72 6.08
specific context aggressive - high p. 4.88 7.46
specific context moderate - high p. 4.98 7.16
specific context Market - high perf. 4.59 7.23
ASE-GI RASE 6.75
Moreover, Table 4 also shows the added value of our approach as the reader can compare our results with the return
on investment (RASE) of the General Index of the Athens Stock Exchange (ASE-GI). According to the results of this
table, the average return of the constructed portfolios for all contexts, except two, achieves higher return than the
market index. The two cases where the constructed portfolios did not beat the market index are the moderate simple
context and moderate-market specific context. This is, maybe, due to the fact that in these two contexts we have an
investor who wishes to earn more without taking into account any amount of risk in relation to the variability which
19
characterizes the conditions of the market during the examined period. This fact makes it very difficult to implement
investment strategies that can help a fund manager outperform a passive investment policy. However, looking at
Table 5, we find that for an investment range of eight years (2000-2007) all contexts are more profitable than the
RASE.
Furthermore, in Table 5 we notice that in some specific contexts the results are more satisfying than the results
obtained by simple contexts, while in others there is little or no difference. This means that by using effective
strategies in the third preference rules layer the decision maker can optimize the combined contexts. Specifically, the
aggressive-high performance specific context provides better results than both the simple contexts aggressive and
high performance (the ones that it combines) and the general context. The moderate-high performance specific
context’s returns on investment are equal to the higher simple context’s returns (high performance) while the
aggressive-market specific context returns are closer to the higher simple context’s returns (market).
Table 5. Resulting Portfolios for all contexts for all years and their returns compared to each other and to RASE. The
values in the grey cells are those that were not successfully forecasted.
Portfolios for contexts 2007 2006 2005 2004 2003 2002 2001 2000 Avg
sim
ple
conte
xts
General -29,79 17,84 25,31 27,44 11,50 23,41 -25,31 -21,19 3,65
Aggressive -31,69 17,64 27,00 29,13 14,56 23,41 -28,65 -21,19 3,78
Moderate -29,43 18,19 24,60 26,17 8,05 23,41 -24,48 -21,19 3,17
Market -26,42 17,47 27,63 29,17 16,50 23,41 -24,69 -21,19 5,24
HighPerfomance -29,77 17,78 21,33 27,63 12,10 27,99 -24,76 -21,31 3,87
specific
conte
xts
Aggressive Market -29,06 17,47 27,63 29,17 15,04 23,41 -26,18 -21,19 4,54
Moderate Market -26,42 18,19 24,60 26,17 16,50 23,41 -24,69 -21,19 4,57
Aggressive HighPerfomance -29,77 17,84 23,90 27,44 11,50 27,99 -24,76 -21,31 4,10
Moderate HighPerfomance -29,77 17,78 21,33 27,63 12,10 27,99 -24,76 -21,31 3,87
Market HighPerfomance -29,77 17,78 21,33 27,63 12,10 27,99 -24,82 -21,31 3,87
RASE (next year) -38,97 15,95 19,95 29,71 20,42 27,38 -33,45 -23,53 2,18
Forecasted as Bear Bull Bull Bull Bear Bull Bear Bull
20
-50
-40
-30
-20
-10
0
10
20
30
40
2000 2001 2002 2003 2004 2005 2006 2007
General
Aggressive
Moderate
Market
HighPerfomance
Aggressive Market
Moderate Market
Aggressive HighPerfomance
Moderate HighPerfomance
Market HighPerfomance
RASE (next year)
Fig. 6: Comparative RoIs of all contexts for each year.
Finally, in Fig. 6, we present the RoI of all contexts separately for each year. This view is also useful, as it shows that
for two years, i.e. 2003 and 2004, RASE was greater than all our contexts RoI performance. This shows that our
application, for the time being, performs better for medium term to long term investments, i.e. those that range over
five years.
All in all, we evaluated our system in two major axes. Firstly, we compared its performance with a previously
developed system [34], which also provides us with the benchmark scenarios (see Table 4). Then, we compared it
with the performance of the Athens General Index (see Table 5). The benchmark data have been retrieved from the
Athens Stock Exchange and such data have been used in the past for evaluating the performance of similar systems.
For example, the systems presented in [5], [20] and [24] have been evaluated in comparison to the S&P 500, an
index of the prices of 500 large cap common stocks actively traded in the two largest American stock markets, the
21
New York Stock Exchange and NASDAQ. Our eight-year performance data demonstrate that our system is useful
for medium to long term investments (as in this time range it always provides a better performance than that of the
ASE-GI).
6 RELATED WORK AND CONCLUSION
The objective of this paper was to present an artificial intelligence based application for the MF portfolio generation
problem that combines two different intelligent methods, argumentation based decision making and a hybrid system
that combines Genetic Algorithms (GA), MultiModel Partitioning (MMP) theory and the Extended Kalman Filter
(EKF).
We described in detail how we developed our argumentation theory and how we combined it with the hybrid system
to determine an important fact for the decision making process, i.e. the status of the financial market in the next
investment period.
We chose argumentation as it allows for decision making using conflicting knowledge, thus different experts can
express their knowledge without needing to be consistent with the opinions of the others. It allows to define nonstatic
priorities between arguments, which means that the priorities of rules can depend on context. Finally, the modularity
of its representation allows for the easy incorporation of views of different experts ([2]). Traditional approaches do
not provide these benefits, on the contrary, they pose specific limitations. Statistical methods need to make a strict
statistical hypothesis not allowing for flexibility ([2]). Multi-criteria analysis methods need significantly more effort
from experts (e.g. Electre-tri, [38]) as they need to define criteria, weights, preference, indifference, veto thresholds,
and profiles, leading, in some cases, to troublesome situations due to time constraints or the reluctance of the
decision makers to actively contribute in a direct interrogation process managed by an expert analyst (see [11] for a
complete comparison between decision support methods). Neural networks have also been used in the past (as in
[25] for example), but they require increased computational effort and are characterized by their inability to provide
meaningful explanations for the produced results ([35]). The reasons for using the selected argumentation framework
([17]) among others ([26]) are a) the well documented Gorgias open source library, and, b) the possibility for
dynamic preference selection in different levels that allowed for the modular conception of the rules (the different
experts could independently express their knowledge).
22
We chose GAs as they comprise a global forecasting technique, in the sense that a single formula is sought that
allows forecasts of future entries in any series generated by the process—starting at any point in time. Local
approximation schemes, on the other hand, deal only with data that lie in a close neighbourhood in the embedding
space, and require separate computations for forecasts in different regions. Obviously, a conventional search for the
actual equation of the data generating process is hopeless, since the dynamics underlying most data, especially
financial series, are far too complicated. The usability of classical regression analysis techniques is also very limited.
During the past decade various nonlinear techniques have been developed to accomplish the task of forecasting. The
nearest neighbour technique ([10]) which requires massive amounts of data embeds the time series in a space of
sufficiently high dimension and then seeks vectors in the historical series that are similar to the one that is to be
predicted. Neural nets also use historical series that are similar to the one that is to be predicted, and the back
propagation of errors, to fine tune a set of parameters, and thereby build a forecasting function that links past values
of the time series with future values ([7]; [37]). Predictions based on radial basis functions ([23]) use global
interpolation techniques and have proven useful in practice when only sparse data was available. Finally, wavelet
analysis—like the Fourier transform—decomposes a time series into its basic components, thus allowing insight into
its fine structure, and then uses a weighted sum of these wavelets for forecasting purposes ([28]).
The developed application allows a decision maker (fund manager) to construct multi-portfolios of MFs under
different, possibly conflicting contexts. Moreover, for medium to long term investments, the returns on investment of
the constructed portfolios are better than those of the General Index of the Athens Stock Exchange, while the best
results are those that involve the forecasting of the financial market.
Our future work will be to develop a new rule base for the problem of determining when to construct a new portfolio
for one of our investors. We will also make the application web-based so that it can get on-line financial data
available from the internet for computing the decision variables and for allowing the investors to insert their profiles
by filling on-line forms. Our aim is to be able to guarantee a better RoI than that of the RASE even for short term
investments.
23
7 REFERENCES
[1] A.V. Adamopoulos, P.A. Anninos, S.D. Likothanassis, G.N. Beligiannis, L.V. Skarlas, E.N. Demiris and
Papadopoulos, P., Evolutionary Self-adaptive Multimodel Prediction Algorithms of the Fetal
Magnetocardiogram, in: Proceedings of the 14th
Int. Conf. on Digital Signal Processing (DSP 2002), Vol. II, 1-
3 July, Santorini, Greece, 2002, pp. 1149-1152.
[2] E.L. Altman, R.A. Eisenbeis, J. Sinkey, Applications of Classification Techniques in Business, Banking, and
Finance, JAI Press, Greenwich, CT, 1981.
[3] L. Amgoud, S. Kaci, An argumentation framework for merging conflicting knowledge bases: The prioritized
case. in: Lecture Notes in Computer Science (LNCS), 3571, Springer-Verlag, 2005, pp. 527-538.
[4] G. N. Beligiannis, L. V. Skarlas and S. D. Likothanassis. A Generic Applied Evolutionary Hybrid Technique for
Adaptive System Modeling and Information Mining, IEEE Signal Processing Magazine 21(3) (2004), 28-38.
[5] P.L. Brockett, A. Charnes and W.W. Cooper, Chance constrained programming approach to empirical analyses
of mutual fund investment strategies, Decision Sciences 23 (1992), 385-403.
[6] S.J. Brown, W.N. Goetzmann, Performance persistence, Journal of Finance 50 (1995), 679-698.
[7] Nonlinear Modeling and Forecasting, SFI Studies in the Sciences of Complexity, M. Casdagli and S. Eubanks
(eds.), Addison-Wesley, New York, 1992.
[8] G. Colson, and M. Zeleny, Uncertain prospects ranking and portfolio analysis under the condition of partial
information, in: Mathematical Systems in Economics 44, Verlag Anton Hain, ed., 1979.
[9] G. Crina, A. Ajith, I. Hisao, eds., Hybrid Evolutionary Algorithms, Studies in Computational Intelligence 75
2007.
[10] R. Doerner, B. Hübinger, and W. Martienssen, Controlling chaos experimentally in systems exhibiting large
effective Lyapunov exponents, Phys. Rev. E 50 (12) (1994), 932-948.
[11] M. Doumpos and C. Zopounidis, Multicriteria Decision Aid Classification Methods, Kluwer Academic
Publishers, Dordrecht, 2002.
[12] E.J. Elton, M.J.Gruber, S. Das, M. Hlavka, Efficiency with costly information: A reinterpretation of evidence
from managed portfolios, The Review of Financial Studies 6(1) (1993), 1-22.
24
[13] C.S. Eun, R. Kolodny, B.G. Resnick, U.S.-based international mutual funds: A performance evaluation, Journal
of Portfolio Management 17 (1991), 88-94.
[14] J.G. Gallo, P.E Swanson, Comparative measures of performance for U.S.-based international equity mutual
funds, Journal of Banking and Finance 20 (1996), 1635-1650.
[15] S. Haykin, Adaptive Filter Theory, Englewood Cliffs, NJ: Prentice-Hall Int., 1991.
[16] R.A. Ippolito, Efficiency with costly information: A study of mutual fund performance, 1965-1984, Quarterly
Journal of Economics 104 (1989), 1-23.
[17] A. Kakas and P. Moraitis, Argumentation based decision making for autonomous agents, in: Proceedings of the
2nd
Int. Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS03), July 14-18, Australia, 2003.
[18] S.K. Katsikas, S.D. Likothanassis, G.N. Beligiannis, K.G. Berketis and D.A. Fotakis, Evolutionary multimodel
partitioning filters: A unified framework, IEEE Trans. Signal Processing 49(10) (2001), 2253-2261.
[19] R.T. LeClair, Discriminant analysis and the classification of mutual funds, Journal of Economics and Business
26(3) (1974), 220-224.
[20] A.L. Loviscek and W.J. Jordan, Stock selection based on Morningstar’s ten-year, five-star general equity mutual
funds, Financial Services Review 9 (2000), 145-157.
[21] H. Markowitz, Portfolio Selection: Efficient Diversification of Investments, Wiley, New York, 1959.
[22] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, 3rd
ed. New York: Springer-
Verlag, 1996.
[23] J. Moody and C. J. Darken, Fast Learning in Networks of Locally-Tuned Processing Units, Neural Computation
1 (1989), 281-294.
[24] D. Morton, E. Popova and I. Popova, Efficient fund of hedge funds construction under downside risk measures,
Journal of Banking & Finance 30(2) (2005), 503–518.
[25] E. Patuwo, M.Y. Hu, M.S. Hung, Two-group classification using neural networks, Decision Sciences 24 (1993)
825-845.
[26] H. Prakken and G. Vreeswijk, Logics for Defeasible Argumentation, in: Handbook of Philosophical Logic,
volume 4, D. Gabbay and F. Guenthner, eds., Kluwer Academic Publishers, 2002, pp. 218-319.
25
[27] I. Rahwan, P. Moraitis and C. Reed, eds., Argumentation in Multi-Agent Systems, Lecture Notes in Artificial
Intelligence, 3366, Springer-Verlag, Berlin, Germany, 2005.
[28] J. B. Ramsey and Z. Zhang, The analysis of foreign exchange data using waveform dictionaries, in: Proceedings
of the 1st
International Conference on High Frequency Data in Finance, Olsen and Associates, Zurich,
Switzerland, 1995.
[29] A. Redman, N. Gullett, H. Manakyan, The performance of global and international mutual funds, Journal of
Financial and Strategic Decisions 13(1) (2000), 75-85.
[30] S. Ross, The Arbitrage Theory of Capital Asset Pricing, Journal of Economic Theory 6 (1976), 341-360.
[31] W.F. Sharpe, Capital asset prices: A theory of market equilibrium under conditions of risk, Journal of Finance
19 (1964), 425-442.
[32] W.F. Sharpe, Mutual fund performance, Journal of Business 39 (1966), 119-138.
[33] K. Simons, Risk-adjusted performance of mutual funds, New England: Economic Review (September/October)
(1998), 33-48.
[34] N. Spanoudakis and K. Pendaraki, A Tool for Portfolio Generation Using an Argumentation Based Decision
Making Framework , in: Proceedings of the annual IEEE International Conference on Tools with Artificial
Intelligence (ICTAI 2007), Patras, Greece, October 29-31, 2007.
[35] V. Subramanian, M.S. Hung, M.Y. Hu, An experimental evaluation of neural networks for classification.
Computers and Operations Research 20(7) (1993), 769-782.
[36] J.L. Treynor, How to rate management of investment funds, Harvard Business Review 43 (1965), 63-75.
[37] A. S. Weigend and N. A. Gershenfeld, Time Series Prediction, SFI Studies in the Sciences of Complexity,
Addison-Wesley, New York, 1994.
[38] W. Yu, ELECTRE TRI: Aspects methodologiques et manuel d’utilisation, Document du Lamsade, No. 74,
Universite de Paris-Dauphine, 1992.